Are insect brains the secret to great AI昆虫大脑是伟大人工智能的秘密吗？中英双语 ted演讲稿.docx

Are insect brains the secret to great AI?昆虫大脑是伟大人工智能的秘密吗？Are insects the key to brain-inspired computing? Neuroscientist Frances S. Chance thinks so. In this buzzy talk, she shares examples of the incredible capabilities of insects - like the dragonfly's deadly accurate hunting skills and the African dung beetle's super strength and shows how untangling the mysterious web of neurons in their tiny brains could lead to breakthroughs in computers, AI and more.昆虫是大脑启发计算的关键吗？神经科学家弗朗西斯s钱斯是这样 认为的。在这场有趣的讲座中，她分享了昆虫不可思议的能力的例子 比如蜻蜓极精确狩猎技能和非洲蜕螂的超强力量并展示了 解开它们微小大脑中神秘的神经元网络如何导致计算机、人工智能等 领域的突破。Creating intelligence on a computer. This has been the Holy Grail for artificial intelligence for quite some time. But how do we get there? 在计算近h创建智能。很长-段时珈姝，这一羁人工智能的 圣杯。但我fi孜口何5哒那里？things as biological brains but does them in the same way as biological brains. This could lead to drones driven by computers the same size of the dragonfly's brain that captures some targets and avoid others. Personally, Pm hoping for a small army of these to defend my backyard from mosquitoes in the summer.但这就割i班始了解大脑尉晌进行基本或原始的H算。计 算，我将其视为更复杂功能的构件，不仅用于拦截，还用于认 知。这些神经的计算方式可能不同于当今计算机Jz存在的任何 东西 这I虹脩勺目标和取 编复制襁沅的横燃代 码。我（门的目标是制衢-种计算机芯片，它不仅可以做与生物 大0解目同的事情，而且可以用与生物大B齿同样0妨式来做区些 事情。这可能会费由计算棚区动的无人机,其大与黜S的大 麟胴，才转理目标并避开其他目阮 雌个人而言，我希 望有一辨辨的无人机在夏天保护我的后院不受蚊?置尤。The GPS on your phone could be replaced by a new navigation device based on dung beetles or ants that could guide you to the straight or the easy path home.你手机上的GPS可能会被一种基于辘良的蚂蚁的新型导航设 备所取代，它可以弓I导你走直路或容易回家的路。And what would the power requirements of these devices be like? As small as it is - Or, sorry - as large as it is, the human brain is estimated to have the same power requirements as a 20-watt light bulb. Imagine if all brain-inspired computers had the same extremely low-power requirements. Your smartphone or your smartwatch probably needs charging every day. Your new brain-inspired device might only need charging every few months, or maybe even every few years.月陷这些设备的功短求是怎样的呢?尽管它很小，或者说，对 不起，尽管它很± ,据估计，人0酬功率需求与20瓦的灯泡 相同。想象一下，如果所有受大脑启发的计算机都具锄目同的 极低咻谦求你0灌能制腌育殍表可能砺雁充 电。你的新大B诟发设备可能只需朝PUI个月,甚到饰充 电一次。The famous physicist, Richard Feynman, once said J What I cannot create, I do not understand/ What I see in insect nervous systems is an opportunity to understand brains through the creation of computers that work as brains do. And creation of these computers will not just be for knowledge. There's potential for real impact on your devices, your vehicles, maybe even artificial intelligences.著名物理物里德费曼曾说：我不能创造的东西，我就不了 解。我在昆虫神经辍中看的g创造与大册-样 工作的计算机来了解大脑的n会。而这些计算机的创造将不仅 仅是了认知。有可育树你的K备、第两甚至是人工智育野生真 国撤响。So next time you see an insect, consider that these tiny brains can lead to remarkable computers. And think of the potential that they offer us for the future.所以，下次你看®卜只昆虫时,想想看，这些微勺大®可以 发展出例的计算机。想®它们辘们断睐提（雌潜力。Thank you.寸谢。So we view ourselves as highly intelligent beings. So it's logical to study our own brains, the substrate of our cognition, for creating artificial intelligence. Imagine if we could replicate how our own brains work on a computer. But now consider the journey that would be required.我（i以为自己是高蟋慧的人。因此，研羯前自己勺大ft我4、蝌口的基础,来创造人工智能是合乎图i的。想象一下，如果我们可以在计算机上复制俄们自己正肉醍如何工作的。但现 在考虑一下所需的过程。The human brain contains 86 billion neurons. Each is constantly communicatmg with thousands of others, and each has individual characteristics of its own. Capturing the human brain on a computer may simply be too big and too complex a problem to tackle with the technology and the knowledge that we have today.年大h鲍含860亿个神经元。每个人都在不断地与成千上万 的人交流，每个人都有自己的特点。在计算机上J赣人脑智慧可 能的确hh太大、太复杂的问题，无法我”将天网玄椅口 知识来解决。I believe that we can capture a bram on a computer, but we have to start smaller. Much smaller.我相信我们可以在计算机H跋智慧，但我们必须从更在）地方 开始。/得多。These insects have three of the most fascinating brains in the world to me. While they do not possess human-level intelligence, each is remarkable at a particular task. Think of them as highly trained specialists.对糠说，这些昆虫有三个世界上最迷人的大0落 虽然讨坏 具备辰水平的智力，但新十都国靛活中裁吐色。将他African dung beetles are really good at rolling large balls in straight lines.非洲屎壳良面学由亶长在直线上滚动大求，Now, if youVe ever made a snowman, you know that rolling a large ball is not easy. Now picture trying to make that snowman when the ball of snow is as big as you are and you!re standing on your head. 如果你曾经堆过雪人，你就知道滚TR求并不容易。现在想象 一下堆雪人当雪爵的一样大时，你倒立着。Sahara desert ants are navigation specialists. They might have to wander a considerable distance to forage for food. But once they do find sustenance, they know how to calculate the straightest path home. 撒0合撞少黜蚁是导专冢他i阿育裹走很密生豺能觅 食。但一旦他fi嵋了食物，他们嬲瞭晌计算回家的最直路 径。And the dragonfly is a hunting specialist. In the wild, dragonflies capture approximately 95 percent of the prey they choose to go after. 而由靛堤狩猎专彖 在野卜，蜻蛔赣了大约95%的它们选 搔豌物。These insects are so good at their specialties that neuroscientists such as myself study them as model systems to understand how animal nervous systems solve particular problems. And in my own research, I study brams to bring these solutions, the best that biology has to offer, to computers.这些昆虫非常M长它们的专业，以至于像我这样0辨经科学家 将它仞乍为陛麴雨腕，以了廨械神经尉胞解夬 特定的问题。磅的研究中，我研究大脑，以将这些生物所能 提供的最好的解;夬方案引入计算机So consider the dragonfly bram. It has only on the order of one million neurons. Now, it's still not easy to unravel a circuit of even one million neurons. But given the choice between trying to tease apart the one-million-neuron brain versus the 86-billion-neuron brain, which would you choose to try first?想一下虫飘的大B理 它只有大约100万个神经元。现在，要解 开f那怕有一鲂个襁航的回路腺楮醺如耨 在尝试施里100万个神经元大0齿和860亿个神经元大S蚯间 做出僻，你会窗圣先尝脚那T ?When studying these smaller insect brains, the immediate goal is not human intelligence. We study these brains for what the msects do well. And in the case of the dragonfly, that!s mterception. So when dragonflies are hunting, they do more than just fly straight at the prey. They fly in such a way that they will intercept it. They aim for where the prey is going to be. Much like a soccer player, running to intercept a pass. To do this correctly, dragonflies need to perform what is known as a coordinate transformation, going from the eye's frame of reference, or what the dragonfly sees, to the body's frame of reference, or how the dragonfly needs to turn its body to intercept. 当研究这域切珀勺昆虫大脚寸，当前的目标不是人类骷僭力。我 微胶这些大脑是为了了解昆虫做得好的地方。就蜻蜓而言, 刃蹴是拦截。因此，当蜻赠地寸，它们所做0勺不仅仅是直接1S 向猎物。它们以这样的方式飞行，以拦截它。它们瞄潮勖将要 至肱的地方。就像足球运动员，跑去拦截但克为了正确地做 至啮一点，蜻蜓需要进行所谓的坐不趟奂,从眼睛的参照系或1 蜓看到的东西，到身体的参照系，或者蜻蜓献如何转动身体进Coordinate transformations are a basic calculation that animals need to perform to mteract with the world. We do them instinctively every time we reach for something. When I reach for an object straight in front of me, my arm takes a very different trajectory than if I turn my head, look at that same object when it is off to one side and reach for it there. In both cases, my eyes see the same image of that object, but my brain is sending my arm on a very different trajectory based on the position of my neck.坐标变：蝇动物与世界互动所需要进行的基本计算，我n我次伸 手拿东西的时候都会本能w嬷些计算。当我伸手去拿我面前 白小物体时,我的手臂的运动轨迹和璐挑看向TS的同T勿 体时完全不同。在这两种情况下，我的瞒青看的都是同T勿 体的图像，但我的大脑m翩醇？的位置将我的手臂墓t一 个非常不同的做。And dragonflies are fast. This means they calculate fast. The latency, or the time it takes for a dragonfly to respond once it sees the prey turn, is about 50 milliseconds. This latency is remarkable. For one thing, it's only half the time of a human eye blink. But for another thing, it suggests that dragonflies capture how to mtercept in only relatively or surprisingly few computational steps.蜻蜓很夬这意口楷他i卅算得很快。延迟，即蜻蜓在看到物 转向后做出反应所需的时诃，大约是50 Wo这种延迟是很 了不起的。一方面这只是幡眨0跳响的T。但另一方面,它 表明蜻蜓仅通函用珀勺或就的极少计算步骤即可体现出如何 亍拦截So in the brain, a computational step is a single neuron or a layer of neurons working in parallel. It takes a single neuron about 10 milliseconds to add up all its inputs and respond. The 50-millisecond response time means that once the dragonfly sees its prey turn, there's only time for maybe four of these computational steps or four layers of neurons, working in sequence, one after the other, to calculate how the dragonfly needs to turn. In other words, if I want to study how the dragonfly does coordinate transformations, the neural circuit that I need to understand, the neural circuit that I need to study, can have at most four layers of neurons. Each layer may have many neurons, but this is a small neural circuit. Small enough that we can identify it and study it with the tools that are available today.所以在大脑中，计算步骤单个神经元或一层襁航并行工 作。单个神经元献大约10薪少潴缩其所有输入相加故 出反应。50翎的响应时间意D精，一期蜓童1它的猎f嫡专 向，可能只有四个计算步骤或醒神经元依次工作的时诃，- 个接T ,来计窠同廷需要如何转向。换句话说，如弄曲研 究虫蠡致小可进行坐标甥奂，我需要了解神经回路，我需要脑 神经回路，最多可以有四层神经元。每一层可能有许多神经 元，但这号/4在神经回路。小至I典i网以用今天的工具来 翊司骑院它。And this is what I'm trying to do. I have built a model of what I believe is the neural circuit that calculates how the dragonfly should turn. And here is the cool result. In the model, dragonflies do coordinate transformations in only one computational step, one layer of neurons. This is something we can test and understand.这就是我要做的。我已箍立了T我认为是计算虫嬴应该姆可 转向的神经回路的摩。这等/很酷&蹄果。在该陛中，蜻 蜓只用ft算步骤，艮f神经元层来做坐信封奂。这是我4、问以溅诳瞳解6勺。In a computer simulation, I can predict the activities of individual neurons while the dragonfly is hunting. For example, here I am predicting the action potentials, or the spikes, that are fired by one of these neurons when the dragonfly sees the prey move. To test the model, my collaborators and I are now comparing these predicted neural responses with responses of neurons recorded in living dragonfly brains. These are ongoing experiments in which we put living dragonflies in virtual reality.在计算机例以中，我可以就则蜻赠守猎时单个神经元的舌动。例 如，我在这里预则当雌看5%昔物移动时，其中T神经元发 射了动作电位撷沌为了测试S个博L物曲效合储现 在正在|等这细而则的神经反应与活体懒g大脑中记录的神经元 反应进行t徽。这些是正在进行的奖佥,我O等活体辘醉在虚 拟贩中。Now, it's not practical to put VR goggles on a dragonfly. So instead, we show movies of moving targets to the dragonfly, while an electrode records activity patterns of individual neurons in the brain. Yeah, he likes the movies. If the responses that we record in the brain match those predicted by the model, we will have identified which neurons are responsible for coordinate transformations. The next step will be to understand the specifics of how these neurons work together to do the calculation.现在，给蜻蜓戴上VR护目镜是襁实的。因此，我们改为向蜻 蜓播放移动目标的电影，同时电极记录封齿中单个神经抽勺活 动最L是的，他戢电影。如果我们在大脑中记录的反应与 匿予五则的反应相匹配，我（门就会确定明陛神经元负责坐标转 换。下T将是了解这些神经元如何协同工作进行计算的细U But this is how we begin to understand how brains do basic or primitive calculations. Calculations that I regard as buildmg blocks for more complex functions, not only for interception but also for cognition.The way that these neurons compute may be different from anything that exists on a computer today. And the goal of this work is to do more than just write code that replicates the activity patterns of neurons. We aim to build a computer chip that not only does the same